U.S. patent application number 16/194941 was filed with the patent office on 2019-05-23 for apparatus and method for ct image denoising based on deep learning.
The applicant listed for this patent is ClariPI Inc., Seoul National University R&DB Foundation. Invention is credited to Jong Hyo KIM, Hyun Sook PARK.
Application Number | 20190156524 16/194941 |
Document ID | / |
Family ID | 66533117 |
Filed Date | 2019-05-23 |
![](/patent/app/20190156524/US20190156524A1-20190523-D00000.png)
![](/patent/app/20190156524/US20190156524A1-20190523-D00001.png)
![](/patent/app/20190156524/US20190156524A1-20190523-D00002.png)
United States Patent
Application |
20190156524 |
Kind Code |
A1 |
PARK; Hyun Sook ; et
al. |
May 23, 2019 |
APPARATUS AND METHOD FOR CT IMAGE DENOISING BASED ON DEEP
LEARNING
Abstract
Provided is a method for CT image denoising based on deep
learning, and the method for CT image denoising based on deep
learning includes: extracting examination information from an input
CT image; selecting at least one deep learning model corresponding
to the examination information from multiple previously trained
deep learning models; and outputting a CT image denoised from the
input CT image by feeding the input CT image into the selected at
least one deep learning model.
Inventors: |
PARK; Hyun Sook; (Seoul,
KR) ; KIM; Jong Hyo; (Seoul, KR) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
ClariPI Inc.
Seoul National University R&DB Foundation |
Seoul
Seoul |
|
KR
KR |
|
|
Family ID: |
66533117 |
Appl. No.: |
16/194941 |
Filed: |
November 19, 2018 |
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06T 2207/10072
20130101; G06T 7/0012 20130101; G06T 11/008 20130101; G06T
2207/20081 20130101; G06T 2207/10081 20130101; G06T 2207/20084
20130101; G06T 11/003 20130101; G06T 5/002 20130101 |
International
Class: |
G06T 11/00 20060101
G06T011/00; G06T 5/00 20060101 G06T005/00; G06T 7/00 20060101
G06T007/00 |
Foreign Application Data
Date |
Code |
Application Number |
Nov 20, 2017 |
KR |
10-2017-0155115 |
Oct 17, 2018 |
KR |
10-2018-0123786 |
Claims
1. A method for CT image denoising based on deep learning,
comprising: extracting examination information from an input CT
image; selecting at least one deep learning model corresponding to
the examination information from multiple previously trained deep
learning models; and outputting a CT image denoised from the input
CT image by feeding the input CT image into the selected at least
one deep learning model.
2. The method for CT image denoising based on deep learning of
claim 1, further comprising: before the extracting process,
generating a second training CT data set to which noises of
multiple predetermined levels are added by applying a CT image
noise simulator to a first training CT data set; extracting
examination information from the second training CT data set and
grouping the second training CT data set into multiple groups
according to a predetermined rule; and generating and training
multiple training-target deep learning models so as to correspond
to the respective groups of the second training CT data set by
groups, wherein in the selecting process, the multiple previously
trained deep learning models are the multiple training-target deep
learning models trained in the training process.
3. The method for CT image denoising based on deep learning of
claim 2, wherein in the generating process, the CT image noise
simulator performs, generating a synthetic sinogram by inputting an
image (original image) of the first training CT data set;
generating a noise-component synthetic sinogram by applying a noise
model having a predetermined level of noise to the synthetic
sinogram; generating a noise-component CT image based on the
generated noise-component synthetic sinogram; and generating a
synthetic high noise CT image by adding the generated
noise-component CT image to the original CT image.
4. The method for CT image denoising based on deep learning of
claim 2, wherein the second training CT data set is composed of a
pair of synthetic high noise CT image set and synthetic
noise-component CT image set obtained on the basis of the first
training CT data set.
5. The method for CT image denoising based on deep learning of
claim 2, wherein the training process includes feeding the
synthetic high noise CT image set for the respective groups of the
second training CT data set into the input of the corresponding
training-target deep learning model and training the
training-target deep learning model repeatedly by groups to
minimize the difference between the synthetic noise-component CT
image and output from the training-target deep learning model, in
order for the multiple training-target deep learning model to have
a function of extracting a noise-component CT image from an input
CT image.
6. The method for CT image denoising based on deep learning of
claim 2, wherein the outputting process includes feeding the input
CT image into the selected at least one deep learning model and
forcing the selected at least one deep learning model to extract a
noise-component CT image from the input CT image, and producing the
denoised CT image by multiplying the extracted noise-component CT
image with a predetermined value and subtracting the result from
the input CT image.
7. An apparatus for CT image denoising based on deep learning,
comprising: an extraction unit configured to extract examination
information from an input CT image; a selection unit configured to
select at least one deep learning model corresponding to the
examination information from multiple previously trained deep
learning models; and an output unit configured to output a CT image
denoised from the input CT image by feeding the input CT image into
the selected at least one deep learning model.
8. The apparatus for CT image denoising based on deep learning of
claim 7, further comprising: a training unit configured to generate
and train multiple training-target deep learning models, wherein
the training unit generates a second training CT data set to which
noises of multiple predetermined levels are added by applying a CT
image noise simulator to a first training CT data set, extracts
examination information from the second training CT data set and
groups the second training CT data set into multiple groups
according to a predetermined rule, and generates and trains
multiple training-target deep learning models so as to correspond
to the respective groups of the second training CT data set by
groups, and the multiple previously trained deep learning models
are the multiple training-target deep learning models trained by
the training unit.
9. The apparatus for CT image denoising based on deep learning of
claim 8, wherein the CT image noise simulator generates a synthetic
sinogram by inputting an image (original image) of the first
training CT data set, generates a noise-component synthetic
sinogram by applying a noise model having a predetermined level of
noise to the synthetic sinogram, generates a noise-component CT
image based on the generated noise-component synthetic sinogram,
and generates a synthetic high noise CT image by adding the
generated noise-component CT image to the original CT image.
10. The apparatus for CT image denoising based on deep learning of
claim 8, wherein the second training CT data set is composed of a
pair of synthetic high noise CT image set and synthetic
noise-component CT image set obtained on the basis of the first
training CT data set.
11. The apparatus for CT image denoising based on deep learning of
claim 8, wherein the training unit feeds synthetic high noise CT
image set for the respective groups of the second training CT data
set into the input of the corresponding training-target deep
learning model and trains the training-target deep learning model
repeatedly by groups to minimize the difference between the
synthetic noise-component CT image and output from the
training-target deep learning model, in order for the multiple
training-target deep learning model to have a function of
extracting a noise-component CT image from an input CT image.
12. The apparatus for CT image denoising based on deep learning of
claim 7, wherein the output unit feeds the input CT image into the
selected at least one deep learning model and forces the selected
at least one deep learning model to extract a noise-component CT
image from the input CT image, and produces the denoised CT image
by multiplying the extracted noise-component CT image with a
predetermined value and subtracting the result from the input CT
image.
13. A non-transitory computer-readable storage medium that stores a
program configured to implement a method according to claim 1 on a
computer.
Description
TECHNICAL FIELD
[0001] The present disclosure relates to an apparatus and method
for CT image denoising based on deep learning.
BACKGROUND
[0002] A computed tomography (CT) scan produces a cross-sectional
image of the body lying in a cylinder-shaped large machine equipped
with an X-ray generator. This enables a user to more clearly see
structures and lesions due to less overlap of structures as
compared to a regular X-ray. Therefore, the CT scan has been widely
used for scans of most of the organs and diseases.
[0003] The quality (resolution, precision) of a CT image is a very
important factor for an accurate diagnose of a lesion. With the
advancement of a CT system, efforts are underway to improve the
quality of a CT image. Examples of the efforts may include a
multi-channel detector technology and a high-speed high-resolution
image reconstruction technology. However, most of the conventional
efforts to improve the quality of a CT image may cause high-dose
radiation exposure and thus cause an increased cancer risk to
patients. Particularly, considering the recent public awareness of
radiation exposure, the efforts to acquire a high-quality
diagnostic image need to be accompanied by efforts to minimize a
radiation dose.
[0004] In an effort to do so, CT scanner manufacturers have
released iterative reconstruction CT systems which enable to
produce noise-reduced CT images at lower radiation dose. However,
it is often not easy to afford the iterative reconstruction CT
systems due to higher expenses than conventional products and
difficulty in disposal of the conventional products. The CT scanner
manufacturers sometimes offer an upgrade program to convert the
legacy CT systems into in the iterative reconstruction CT systems.
However, those upgrade programs are not applicable to all legacy CT
systems due to technical problems and high cost. The present
invention was devised to provide a solution to these problems. That
is, the development of more effective technology capable of
acquiring a high-quality diagnostic image (CT image) while
minimizing a radiation dose is needed.
[0005] Meanwhile, deep learning technology refers to a new
computing technology that can achieve a specific goal by using a
general-purpose program code which can be trained according to a
given learning data set instead of coding a series of detailed
programs and has been acknowledged with its great performance in
various image processing fields.
[0006] However, in order for a deep learning model to provide
desirable performance, it is required to secure enough training
data sets. Further, in order for the deep learning model to be
trained about data in a predefined range and to operate in response
to the data in the predefined range during actual operation, a
method for segmenting data is needed. In order to apply deep
learning to medical images for which safety is particularly
important, it can be more important to develop an effective deep
learning technology that fulfills this prerequisite.
[0007] The background technology of the present disclosure is
disclosed in Korean Patent Laid-open Publication No.
10-2014-0130784.
SUMMARY
[0008] In view of the foregoing, the present disclosure provides an
apparatus and method for CT image denoising based on deep learning,
which can acquire a high-quality diagnostic image (CT image) while
minimizing required radiation dose.
[0009] Further, the present disclosure provides a training
(learning) method for a deep learning model for an apparatus for CT
image denoising based on deep learning, which enables deep learning
to be more effectively applied to medical images for which safety
is particularly important.
[0010] Furthermore, the present disclosure provides an apparatus
and method for CT image denoising based on deep learning, which
uses the advantages of deep learning and can ensure the performance
of deep learning and the assurance of safety, and also provides a
method for training a deep learning model using the same.
[0011] However, problems to be solved by the present disclosure are
not limited to the above-described problems. There may be other
problems to be solved by the present disclosure.
[0012] According to an aspect of the present disclosure, there is
provided a method for CT image denoising based on deep learning,
including: extracting examination information from an input CT
image; selecting at least one deep learning model corresponding to
the examination information from multiple previously trained deep
learning models; and outputting a CT image denoised from the input
CT image by feeding the input CT image into the selected at least
one deep learning model.
[0013] According to an embodiment of the present disclosure, the
method for CT image denoising based on deep learning further
includes, before the extracting process, generating a second
training CT data set to which noises of multiple predetermined
levels are added by applying a CT image noise simulator to a first
training CT data set; extracting examination information from the
second training CT data set and grouping the second training CT
data set into multiple groups according to a predetermined rule;
and generating and training multiple training-target deep learning
models corresponding to the respective groups of the second
training CT data set. In the selecting process, the multiple
previously trained deep learning models may be the multiple
training-target deep learning models trained in the training
process.
[0014] Further, in the generating process, the CT image noise
simulator may perform: generating a synthetic sinogram by inputting
an image (original image) of the first training CT data set;
generating a noise-component synthetic sinogram by applying a noise
model having a predetermined level of noise to the synthetic
sinogram; generating a noise-component CT image based on the
generated noise-component synthetic sinogram; and generating a
synthetic high noise CT image by adding the generated
noise-component CT image to the original CT image.
[0015] Furthermore, the second training CT data set may be composed
of a pair of synthetic high noise CT image set and synthetic
noise-component CT image set obtained on the basis of the first
training CT data set.
[0016] Moreover, the training process may include feeding the
synthetic high noise CT image set for the respective groups of the
second training CT data set into the input of the corresponding
training-target deep learning model, and training the
training-target deep learning model repeatedly by groups to
minimize the difference between the synthetic noise-component CT
image and the output from the training-target deep learning model
in order for the training-target deep learning model to have a
function of extracting a noise-component CT image from an input CT
image.
[0017] Further, the outputting process may include feeding the
input CT image into the selected at least one deep learning model
and forcing the selected at least one deep learning model to
extract a noise-component CT image from the input CT image, and
producing the denoised CT image by multiplying the extracted
noise-component CT image with a predetermined value and subtracting
the result from the input CT image.
[0018] Meanwhile, according to another aspect of the present
disclosure, there is provided an apparatus for CT image denoising
based on deep learning, including: an extraction unit configured to
extract examination information from an input CT image; a selection
unit configured to select at least one deep learning model
corresponding to the examination information from multiple
previously trained deep learning models; and an output unit
configured to output a CT image denoised from the input CT image by
feeding the input CT image into the selected at least one deep
learning model.
[0019] According to an embodiment of the present disclosure, the
apparatus for CT image denoising based on deep learning further
includes a training unit configured to generate and train multiple
training-target deep learning models, and the training unit may
generate a second training CT data set to which noises of multiple
predetermined levels are added by applying a CT image noise
simulator to a first training CT data set, extract examination
information from the second training CT data set and group the
second training CT data set into multiple groups according to a
predetermined rule, and generate and train multiple training-target
deep learning models corresponding to the respective groups of the
second training CT data set. The multiple previously trained deep
learning models may be the multiple training-target deep learning
models trained by the training unit.
[0020] Further, the CT image noise simulator may generate a
synthetic sinogram by inputting an image (original image) of the
first training CT data set, generate a noise-component synthetic
sinogram by applying a noise model having a predetermined level of
noise to the synthetic sinogram, generate a noise-component CT
image based on the generated noise-component synthetic sinogram,
and generate a synthetic high noise CT image by adding the
generated noise-component CT image to the original CT image.
[0021] Furthermore, the second training CT data set may be composed
of a pair of synthetic high noise CT image set and synthetic
noise-component CT image set obtained on the basis of the first
training CT data set.
[0022] Moreover, the training unit may feed synthetic high noise CT
image set for the respective groups of the second training CT data
into the input of the corresponding training-target deep learning
model and train the training-target deep learning model repeatedly
by groups to minimize the difference between the synthetic
noise-component CT image and the output from the training-target
deep learning model in order for the training-target deep learning
model to have a function of extracting a noise-component CT image
from an input CT image.
[0023] Besides, the output unit may feed the input CT image into
the selected at least one deep learning model and force the
selected at least one deep learning model to extract a
noise-component CT image from the input CT image, and produce the
denoised CT image by multiplying the extracted noise-component CT
image with a predetermined value and subtracting the result from
the input CT image.
[0024] The above-described aspects are provided by way of
illustration only and should not be construed as liming the present
disclosure. Besides the above-described embodiments, there may be
additional embodiments described in the accompanying drawings and
the detailed description.
[0025] According to the above-described aspects of the present
disclosure, CT image denoising is performed based on deep learning,
and, thus, it is possible to acquire a high-quality diagnostic
image (CT image) while minimizing required radiation dose.
[0026] According to the above-described aspects of the present
disclosure, multiple deep learning models (multiple training-target
deep learning models) are trained according to examination
information of a CT image, which enables deep learning to be more
effectively applied to medical images for which safety is
particularly important.
[0027] According to the above-described aspects of the present
disclosure, it is possible to provide an apparatus and method for
CT image denoising based on deep learning, which uses the
advantages of deep learning and can ensure the performance of deep
learning and the assurance of safety, and a training (learning)
method for deep learning model using the same.
[0028] However, effects to be obtained by the present disclosure
are not limited to the above-described effects. There may be other
effects to be obtained by the present disclosure.
BRIEF DESCRIPTION OF THE DRAWINGS
[0029] In the detailed description that follows, embodiments are
described as illustrations only since various changes and
modifications will become apparent to those skilled in the art from
the following detailed description. The use of the same reference
numbers in different figures indicates similar or identical
items.
[0030] FIG. 1 is a diagram showing a schematic configuration of an
apparatus for CT image denoising based on deep learning according
to an embodiment of the present disclosure.
[0031] FIG. 2 is a diagram showing a schematic operation flow of a
method for CT image denoising based on deep learning through an
apparatus for CT image denoising based on deep learning according
to an embodiment of the present disclosure.
[0032] FIG. 3 is a diagram showing a schematic operation flow
example of a method for training a deep learning model for CT image
denoising based on deep learning through an apparatus for CT image
denoising based on deep learning according to an embodiment of the
present disclosure.
DETAILED DESCRIPTION
[0033] Hereinafter, embodiments of the present disclosure will be
described in detail with reference to the accompanying drawings so
that the present disclosure may be readily implemented by those
skilled in the art. However, it is to be noted that the present
disclosure is not limited to the embodiments but can be embodied in
various other ways. In drawings, parts irrelevant to the
description are omitted for the simplicity of explanation, and like
reference numerals denote like parts through the whole
document.
[0034] Through the whole document, the term "connected to" or
"coupled to" that is used to designate a connection or coupling of
one element to another element includes both a case that an element
is "directly connected or coupled to" another element and a case
that an element is "electronically connected or coupled to" another
element via still another element.
[0035] Through the whole document, the terms "on", "above", "on an
upper end", "below", "under", and "on a lower end" that are used to
designate a position of one element with respect to another element
include both a case that the one element is adjacent to the other
element and a case that any other element exists between these two
elements.
[0036] Further, through the whole document, the term "comprises or
includes" and/or "comprising or including" used in the document
means that one or more other components, steps, operation and/or
existence or addition of elements are not excluded in addition to
the described components, steps, operation and/or elements unless
context dictates otherwise.
[0037] FIG. 1 is a diagram showing a schematic configuration of an
apparatus for CT image denoising based on deep learning according
to an embodiment of the present disclosure. Hereinafter, an
apparatus for CT image denoising based on deep learning 10
according to an embodiment of the present disclosure will be
referred to as the present apparatus 10 for convenience in
explanation.
[0038] Referring to FIG. 1, the present apparatus 10 may include an
extraction unit 11, a selection unit 12, an output unit 13, and a
training unit 14.
[0039] The extraction unit 11 may extract examination information
from an input CT image. Particularly, the extraction unit 11 may
extract the examination information from header information
included in the input CT image. The examination information may
include information about an examination area of the CT image and
scan property information.
[0040] Herein, the information about an examination area may be,
for example, information about an organ area. That is, the
information about an examination area may refer to information
about a significant human organ area of interest and may include
information about organ areas such as head, chest, abdomen, etc.
Further, the scan property information may refer to information
about CT scan variables which affect noise characteristics of the
CT image and may include, for example, information about
reconstructed kernel, tube voltage, tube current, slice thickness,
etc.
[0041] The input CT image may be an image obtained using Computed
Tomography (CT), Magnetic Resonance Imaging (MRI), Positron
Emission Tomography (PET), etc., but may not be limited
thereto.
[0042] The selection unit 12 may select at least one deep learning
model corresponding to the examination information extracted by the
extraction unit 11 from multiple previously trained deep learning
models.
[0043] The selection unit 12 may select at least one deep learning
model (one or more deep learning models) corresponding to the
examination information extracted by the extraction unit 11 from
the multiple previously trained deep learning models by applying a
predetermined rule (previously set rule) to the examination
information extracted by the extraction unit 11.
[0044] Accordingly, the present apparatus 10 may include multiple
deep learning models. Herein, each of the multiple deep learning
models is a deep learning model previously trained by the training
unit 14 and particularly refers to a deep learning model which has
been previously trained to output a CT image (denoised CT image)
denoised from an input CT image fed into the deep learning
model.
[0045] According to another embodiment, each of the multiple deep
learning models included in the present apparatus 10 may refer to a
deep learning model which has been previously trained to extract a
noise-component CT image from an input CT image.
[0046] The multiple deep learning models considered by the present
apparatus 10 may be generated by the training unit 14. The training
unit 14 may generate multiple deep learning models according to
examination information. Particularly, the training unit 14 may
generate multiple deep learning models according to combinations of
examination area information and scan property information included
in the examination information of the second training CT image
set.
[0047] In other words, the present apparatus 10 may include
multiple deep learning models which have been generated according
to the examination information (examination area information and
scan property information). Each of the multiple deep learning
models which have been generated according to combinations of
examination area information and scan property information may be
trained by the training unit 14 using the corresponding second
training CT data set (i.e., CT data sets for deep learning) which
have been grouped according to combinations of examination area
information and scan property information.
[0048] In the present disclosure, deep learning models which are
not yet trained by the training unit 14 may be referred to as
training-target deep learning models. Accordingly, after training,
training-target deep learning models will be referred to as
previously trained deep learning models. In other words, the
multiple previously trained deep learning models considered by the
selection unit 12 may refer to the multiple training-target deep
learning models which have been trained by the training unit 14.
Training of the training-target deep learning models will be
described later in more detail.
[0049] The selection unit 12 may select at least one deep learning
model corresponding to the examination information (examination
area information and scan property information extracted by the
extraction unit) extracted by the extraction unit 11 from the
multiple previously trained deep learning models which have been
generated according to the examination information (combinations of
examination area information and scan property information) and
trained.
[0050] The selection unit 12 may select at least one deep learning
model for an input CT image using a predetermined rule from the
multiple previously trained deep learning models based on the
examination information extracted by the extraction unit 11.
[0051] For example, we assume an input CT image is an abdominal CT
image which is of the first examination area type (e.g., abdomen)
and of the first scan property type (e.g., scan property of low
dose and thick slice). In this case, the selection unit 12 may
select a single deep learning model (e.g., first deep learning
model) corresponding to the first examination area type (e.g.,
abdomen) and the first scan property type which has been trained
with a group of second training data set where the examination area
type is abdomen with scan property type of low dose and thick slice
from the multiple previously trained deep learning models.
[0052] For another example, we assume an input CT image is an
abdominal CT image which is of the first examination area type
(e.g., abdomen) and of the second scan property type (e.g., scan
properties of ultra-low dose and thin slice). In this case, the
selection unit 12 may select a single deep learning model (e.g.,
second deep learning model) corresponding to the first examination
area type and the second scan property type which has been trained
with a group of second training data set where the examination area
type is abdomen with scan property type of ultra-low dose and thin
slice from the multiple previously trained deep learning
models.
[0053] For yet another example, we assume an input CT image is a
chest CT image which is of a second examination area type (e.g.,
chest) and of the third scan property type for the first section
(e.g., applying the reconstructed kernel having high sharpening
property to the lung section) and of the fourth scan property to
the second section (e.g., applying the reconstructed kernel having
low sharpening to the soft tissue section). In this case, the
selection unit 12 may select two deep learning models including a
third deep learning model corresponding to the second examination
area type and the third scan property type and a fourth deep
learning model corresponding to the second examination area type
and the fourth scan property type from the multiple previously
trained deep learning models. To be specific, the selection unit 12
may select two deep learning models including the third deep
learning model which has been trained with a group of the second
training data set where the examination area type is chest and the
scan property type is the reconstructed kernel having high
sharpening property and the fourth deep learning model where the
examination area type is chest and the scan property type is the
reconstructed kernel having low sharpening property from the
multiple previously trained deep learning models.
[0054] As such, the selection unit 12 may select one or more deep
learning models (i.e., at least one deep learning model) from the
multiple previously trained deep learning models based on
examination information extracted by the extraction unit 11 in
consideration of the combination of examination area information
and scan property information included in the extracted examination
information.
[0055] The output unit 13 may output a CT image denoised from an
input CT image by feeding the input CT image into the at least one
deep learning model selected by the selection unit 12. That is, the
output unit 13 may feed the input CT image into the at least one
deep learning model selected by the selection unit 12 and then
output a CT image denoised from the input CT image by taking output
of the selected at least one deep learning model.
[0056] The output unit 13 may feed the input CT image into the at
least one deep learning model selected by the selection unit 12 and
thus force the selected at least one deep learning model to extract
a noise-component CT image from the input CT image, and produce a
denoised CT image by taking the extracted noise from the selected
at least one deep learning model and multiplying with a
predetermined value (a predetermined ratio), and then subtracting
the result from the input CT image.
[0057] In an embodiment, it can be assumed that the selection unit
12 selects a single (one) deep learning model. In this case, the
output unit 13 may feed the input CT image into the single deep
learning model selected by the selection unit 12 and thus force the
selected single deep learning model to extract a noise-component CT
image from the input CT image, and produce a denoised CT image by
multiplying the extracted noise-component CT image with a
predetermined ratio and subtracting the result from the input CT
image.
[0058] In another embodiment, it can be assumed that the selection
unit 12 selects multiple deep learning models. In this case, the
output unit 13 may feed the input CT image into the multiple deep
learning models selected by the selection unit 12 and thus force
the selected deep learning models to extract noise-component CT
images from the input CT image, and produce a denoised CT image by
mixing the multiple noise-component CT images extracted by the
respective multiple deep learning models according to a
predetermined rule followed by multiplying with a predetermined
ratio and subtracting the result from the input CT image.
[0059] The training unit 14 may generate and train multiple
training-target deep learning models before the extraction unit 11
extracts examination information from an input CT image.
[0060] Herein, the training unit 14 may generate the second
training CT data set to which noises of multiple predetermined
levels are added by applying a CT image noise simulator to the
first training CT data set, extract examination information from
the second training CT data set and group the second training CT
data set into multiple groups according to a predetermined rule,
and generate and train multiple training-target deep learning
models corresponding to the respective groups of the second
training CT data set. The training unit 14 may train each of the
generated multiple training-target deep learning models.
[0061] Herein, once training was completed, those generated
multiple training-target deep learning models trained by the
training unit 14 may refer to the multiple previously trained deep
learning models considered by the selection unit 12.
[0062] Further, the CT image noise simulator considered when the
training unit 14 generates the second training CT data set may
generate a synthetic sinogram by inputting an image (original
image) of the first training CT data set, generate a
noise-component synthetic sinogram by applying a noise model having
a predetermined level (desired level) of noise to the synthetic
sinogram, generate a noise-component CT image based on the
generated noise-component synthetic sinogram, and generate a
synthetic high noise CT image by adding the generated
noise-component CT image to the original CT image.
[0063] Furthermore, the second training CT data set considered by
the training unit 14 may be composed of a pair of synthetic high
noise CT image set and synthetic noise-component CT image set
obtained on the basis of the first training CT data set.
[0064] Accordingly, the training unit 14 may train the
training-target deep learning models using a pair of synthetic high
noise CT image set and synthetic noise-component CT image set
obtained by applying the CT image noise simulator to the first
training CT data set.
[0065] Further, the training unit 14 may feed the synthetic high
noise CT image set for the respective groups of the second CT data
set into the input of the corresponding training-target deep
learning model and train the training-target deep learning model
repeatedly by groups to minimize the difference between the
synthetic noise-component CT image and the output from the
training-target deep learning model in order for the
training-target deep learning model to have a function of
extracting a noise-component CT image from an input CT image.
[0066] Hereinafter, the training unit 14 will be described in more
detail.
[0067] The training unit 14 may train a training-target deep
learning model using a training CT data set (i.e., CT data set for
deep learning training) which has been generated through the CT
image noise simulator (the training process takes place before the
extraction unit 11 extracts examination information from an input
CT image). Herein, the training-target deep learning model may
refer to at least one deep learning model.
[0068] Before training the training-target deep learning model, the
training unit 14 may generate a second training CT data set with
noises of multiple predetermined levels (i.e., to which noises of
multiple predetermined levels are added) by applying each CT image
of the first training CT data set as input into the CT image noise
simulator. Then, the training unit 14 may train the training-target
deep learning model using the generated second training CT data
set.
[0069] Herein, the training unit 14 may train training-target deep
learning models corresponding to deep learning models included in
the present apparatus 10 using the generated second training CT
data set in order for the deep learning models to output a low
noise CT image denoised from a high noise CT image fed into the
deep learning models. In other words, through training of the
training-target deep learning models by the training unit 14, each
of the deep learning models included in the present apparatus 10
may be trained to output a denoised CT image when a high noise CT
image is fed.
[0070] After generating the second training CT data set to which
noises of multiple predetermined levels are added, the training
unit 14 may extract examination information (including examination
area information and scan property information) from the second
training CT data set and group the second training CT data set into
multiple groups according to a predetermined rule. Accordingly, the
second training CT data set after grouping according to the
predetermined rule may result in groups of second training CT data
according to examination areas and groups of the second training CT
data according to scan properties.
[0071] To be specific, the training unit 14 may generate the second
training CT data set for the respective noise levels when
generating training CT data set. In other words, the training unit
14 may generate multiple groups of second training CT data set
(i.e., CT data set for deep learning) for the respective
predetermined noise levels when generating the second training CT
data set to which noises of multiple predetermined levels are
added.
[0072] Then, the training unit 14 may extract examination
information including examination area information and scan
property information from the second CT data set which have been
previously generated for the respective noise levels. Then, the
training unit 14 may group the second training CT data set into
multiple groups according to a predetermined rule based on the
extracted examination information. Herein, the predetermined rule
may refer to a rule for classifying CT data (CT images) of the
second CT data set (which have been generated for the respective
noise levels) into predetermined examination area types and/or
predetermined scan property types.
[0073] Accordingly, the training unit 14 may classify the second CT
data set generated for the respective noise levels into multiple
examination area types and/or scan property types and group into
multiple groups. That is, the multiple groups of the second
training CT data set may include the groups of the second training
CT data set classified according to the respective examination
areas and the groups of the second training CT data set classified
according to the respective scan properties.
[0074] Then, the training unit 14 may generate multiple
training-target deep learning models corresponding to the multiple
groups of the second training CT data sets. That is, according to
the present disclosure, the multiple training-target deep learning
models generated by the training unit 14 may include multiple
training-target deep learning models generated corresponding to the
respective examination areas used to classify the second CT data
set and multiple training-target deep learning models generated
corresponding to the respective scan properties used to classify
the second CT data set.
[0075] Then, the training unit 14 may train the generated multiple
training-target deep learning models with the second training CT
data set which have been generated corresponding thereto.
[0076] That is, the training unit 14 may train the multiple
training-target deep learning models generated corresponding to the
respective examination areas using the groups of the second
training CT data set classified according to the respective
examination areas. Further, the training unit 14 may train the
multiple training-target deep learning models generated
corresponding to the respective scan properties using the second
training CT data set classified according to the respective scan
properties.
[0077] In other words, the training unit 14 may train (teach) each
of the multiple training-target deep learning models generated for
the respective examination areas using each corresponding group of
the second training CT data set classified according to the
respective examination areas. Further, the training unit 14 may
train (teach) each of the multiple training-target deep learning
model generated for the respective scan properties using each
corresponding group of the second training CT data set classified
according to the respective scan properties.
[0078] Accordingly, the training unit 14 may train (teach) each of
the multiple training-target deep learning models (multiple deep
learning models) generated corresponding to the respective
examination areas and the respective scan properties and included
in the present apparatus 10 using each group of the second training
CT data set classified according to the respective examination
areas and the respective scan properties in order for each of the
multiple training-target deep learning models to output a denoised
CT image for a high noise CT.
[0079] After generating the multiple training-target deep learning
models corresponding to the multiple groups of the second CT data
set, the training unit 14 may train the multiple training-target
deep learning models using the multiple groups of the second
training CT data set which have been generated corresponding
thereto.
[0080] Through training by the training unit 14, each of the
multiple deep learning models included in the present apparatus 10
may be trained (may learn) to output a low noise CT image denoised
from a high noise CT image when a high noise CT image is fed into
the deep learning model.
[0081] In other words, the training unit 14 may take an original CT
image set (i.e., first training CT data set) as input and generate
a new training CT data set for respective noise levels (i.e.,
second training CT data set for respective noise levels) by using
the CT image noise simulator.
[0082] Then, for example, the training unit 14 may group the second
training CT data set generated for the respective noise levels into
multiple groups. That is, the training unit 14 may generate
multiple groups of the second training CT data set including the
multiple groups of the second training CT data set grouped
according to the respective scan properties based on the second
training CT data set generated for the respective noise levels.
[0083] Then, for example, the training unit 14 may generate, as
training-target deep learning models included in the present
apparatus 10, multiple training-target deep learning models for the
respective scan properties so as to correspond to the multiple
groups of the second CT data set grouped according to the
respective scan properties.
[0084] Then, the training unit 14 may train the multiple
training-target deep learning models generated for the respective
scan properties using the multiple groups of the second training CT
data set grouped according to the respective scan properties.
Herein, the training unit 14 may train (teach) each of the multiple
training-target deep learning models generated for the respective
scan properties in order for the deep learning model
(training-target deep learning model) corresponding to the scan
properties of an input CT image to output a denoised CT image when
a high noise CT image with the scan property of same group is fed
to input.
[0085] Accordingly, when an input CT image is fed to the present
apparatus 10, the selection unit 12 may select a deep learning
model according to the scan property type of the input CT image to
operate the deep learning model trained with the corresponding
group of the second training CT data set having the same scan
property type of the input CT image from multiple previously
trained deep learning models and then, the output unit 13 may
output the denoised input CT image using the selected deep learning
model.
[0086] Further, the training unit 14 may take the first training CT
data set as input and generate a second training CT data set with
noises of multiple predetermined levels by using the CT image noise
simulator.
[0087] Then, for example, the training unit 14 may group the second
training CT data set generated for the respective noise levels into
multiple groups according to the respective examination area types.
That is, the training unit 14 may generate multiple groups of the
second training CT data set including the multiple groups of the
second training CT data set grouped according to the respective
examination area types based on the second training CT data set
generated for the respective noise levels.
[0088] Then, for example, the training unit 14 may generate, as
training-target deep learning models included in the present
apparatus 10, multiple training-target deep learning models for the
respective examination area type so as to correspond to the
multiple groups of the second CT data set grouped according to the
respective examination area types.
[0089] Then, the training unit 14 may train the multiple
training-target deep learning models generated for the respective
examination area types using the multiple groups of the second
training CT data set grouped according to the respective
examination area types. Herein, the training unit 14 may train
(teach) each of the multiple training-target deep learning models
generated for the respective examination area types in order for a
deep learning model (training-target deep learning model)
corresponding to the examination area type of an input CT image to
output a denoised CT image when a high noise CT image is fed to
input.
[0090] Accordingly, when an input CT image is fed to the present
apparatus 10, the selection unit 12 may select a deep learning
model corresponding to the examination area type of the input CT
image to operate the deep learning model trained with the
corresponding group of the second training CT data having the same
examination area type of the input CT image from the multiple
previously trained deep learning models and then, the output unit
13 may output the denoised input CT image using the selected deep
learning model.
[0091] That is, the training unit 14 may generate a second training
CT data set grouped according to the respective examination areas
and/or respective scan properties and generate and train
training-target deep learning models for the respective examination
areas and/or respective scan properties.
[0092] Further, the second training CT data set generated by the
training unit 14 may include at least one pair of image set among
the original CT image set (i.e., the first training CT data set),
synthetic high noise CT image set, and synthetic noise-component CT
image set. In other words, the training unit 14 may generate the
second training CT data set including at least one pair of image
set among the original CT image set, the synthetic high noise CT
image set, and the synthetic noise-component CT image set by using
the original CT image set as input into the CT image noise
simulator.
[0093] The second training CT data set considered by the training
unit 14 may be composed of a pair of the synthetic high noise CT
image set and the synthetic noise-component CT image set derived
from the first training CT data set.
[0094] The training unit 14 may feed the synthetic high noise CT
image set for the respective groups of the second training CT data
set as input into a corresponding training-target deep learning
model and train the training-target deep learning model repeatedly
by groups to minimize the difference between the synthetic
noise-component CT image set and the output from the
training-target deep learning model.
[0095] That is, the training unit 14 may feed the synthetic high
noise CT image set for the respective groups of the second training
CT data set as input into the corresponding training-target deep
learning model and train the training-target deep learning model
repeatedly by groups to minimize the difference between the
synthetic noise-component CT image and the output from the
training-target deep learning model in order for each of the
multiple training-target deep learning models to have a function of
extracting a noise-component CT image from an input CT image.
[0096] In other words, the training unit 14 may feed each synthetic
high noise CT image for the respective groups of the second
training CT data set as an input into the corresponding
training-target deep learning model sequentially and train the
training-target deep learning model repeatedly by groups to
minimize the difference between the synthetic noise-component CT
image and the output from the training-target deep learning model
in order for each of the training-target deep learning model to
have a function of extracting a noise component CT image from an
input CT image.
[0097] Further, the CT image noise simulator used for training by
the training unit 14 may generate a synthetic sinogram using an
image (original image) of the first training CT data set as input.
Furthermore, the CT image noise simulator may generate a
noise-component synthetic sinogram by applying a noise model having
a predetermined level (desired level) of noise to the generated
synthetic sinogram. Moreover, the CT image noise simulator may
generate a noise-component CT image based on the generated
noise-component synthetic sinogram. Besides, the CT image noise
simulator may generate a synthetic high nose CT image by adding the
generated noise-component CT image to the original CT image.
[0098] As a specific example, the CT image noise simulator may
generate a synthetic sinogram using an image (original image) of
the first training CT data set as input. Then, the CT image noise
simulator may generate a noise-component synthetic sinogram by
applying the CT physical principle (i.e., physical principle of CT
scan) as a noise model having a predetermined level of noise to the
generated synthetic sinogram. Then, the CT image noise simulator
may generate a noise-component CT image based on the generated
noise-component synthetic sinogram. Then, the CT image noise
simulator may generate a synthetic high nose CT image to be
included in the second training CT data set for training by adding
the generated noise-component CT image to the original CT
image.
[0099] Herein, when generating a synthetic sinogram, the CT image
noise simulator may determine an attenuation coefficient for each
pixel of an original CT image, distance information between an
X-ray tube focus and a detector, and distance information between
the X-ray tube focus and a subject corresponding to an input CT
image based on examination information of the original CT
image.
[0100] Then, the CT image noise simulator may generate a synthetic
sinogram based on the determined attenuation coefficient for each
pixel, distance information between the X-ray tube focus and the
detector, and distance information between the X-ray tube focus and
the subject. To be specific, the CT image noise simulator may
generate a synthetic sinogram by performing a projection operation
for each rotation angle based on the determined attenuation
coefficient for each pixel, distance information between the X-ray
tube focus and the detector, and distance information between the
X-ray tube focus and the subject.
[0101] As such, the present apparatus 10 may train each of multiple
deep learning models (i.e., multiple training-target deep learning
models) through the training unit 14 in order for each of the
multiple deep learning models included in the present apparatus 10
to output a low noise CT image denoised from an input CT image fed
into the deep learning model. Based on these multiple previously
trained deep learning models (multiple training-target deep
learning models trained by the training unit), the present
apparatus 10 may select a deep learning model corresponding to
examination information extracted from the input CT image from the
multiple previously trained deep learning models. Then, the present
apparatus 10 may feed the input CT image into the selected deep
learning model, and, thus, the selected deep learning model can
output a denoised CT image (i.e., low noise CT image) denoised from
the input CT image.
[0102] The present disclosure provides the apparatus for CT image
denoising based on deep learning (the present apparatus 10) that
outputs a denoised CT image of high quality (low noise CT image)
from an input CT image (e.g., low dose input CT image) of a low
quality based on multiple previously trained deep learning models.
The denoised input CT image output by the present apparatus 10 may
have a similar or higher quality compared to a high dosage CT
image.
[0103] Further, the present apparatus 10 provides a training
(learning) method for deep learning model (training-target deep
learning model) to output a denoised high quality CT image from a
low quality input CT image. The denoised input CT image output by
the trained deep learning model (i.e., previously trained deep
learning model) may have a similar or higher quality than a high
dosage CT image.
[0104] In the present apparatus 10, the extraction unit 11 may
extract examination area information and scan property information
as examination information from an input CT image using examination
information included in the input CT image. Then, the present
apparatus 10 may select a deep learning model (i.e., deep learning
model previously trained to provide a predetermined level of
performance) corresponding to the extracted examination information
from multiple previously trained deep learning models and output
(obtain) a denoised input CT image through the selected deep
learning model. In the present apparatus 10, the training unit 14
may train each of multiple training-target deep learning models in
order for each of the multiple deep learning models to provide a
predetermined level of performance (a level at which the difference
between the synthetic noise-component CT image and output from the
training-target deep model is reached).
[0105] Further, the present apparatus 10 may receive the first
training CT data set as input and generate a new training CT data
set (second training CT data set) including a synthetic high noise
CT image set by using the CT noise simulator and may train
training-target deep learning models included in the present
apparatus 10 using the generated second training CT data set. The
training-target deep learning models included in the present
apparatus 10 can become previously trained deep learning models
through training using the second training CT data set, and these
previously trained deep learning models may refer to the deep
learning models previously trained to output a denoised CT image
(low noise CT image) from a high noise CT image.
[0106] Furthermore, the present apparatus 10 trains the
training-target deep learning models using the second training CT
data set generated by using the CT noise simulator, and, thus, when
a high noise CT image (input CT image) actually obtained by low
dose scanning is fed into the present apparatus 10 (i.e., input
into at least one of the multiple previously trained deep learning
models included in the present apparatus), the present apparatus 10
can more effectively output a low noise CT image denoised from the
high noise CT image.
[0107] Each of the multiple deep learning models generated for the
respective examination area types and respective scan property
types and included in the present apparatus 10 can be trained by
the training unit 14 to output a CT image denoised from an input CT
image.
[0108] The present apparatus 10 can be referred to as an apparatus
for training a deep learning model (training-target deep learning
model) for an apparatus for CT image denoising based on deep
learning as well as an apparatus for CT image denoising based on
deep learning. The present disclosure can provide a method for
training a deep learning model for CT image denoising based on deep
learning as well as a method for CT image denoising based on deep
learning through the present apparatus 10.
[0109] Hereinafter, the flow of the operations of the present
disclosure will be described briefly based on the above
details.
[0110] FIG. 2 is a diagram showing a schematic operation flow of a
method for CT image denoising based on deep learning through an
apparatus for CT image denoising based on deep learning according
to an embodiment of the present disclosure.
[0111] The method for CT image denoising based on deep learning
illustrated in FIG. 2 can be performed by the above-described
apparatus for CT image denoising based on deep learning (the
present apparatus 10). Accordingly, the descriptions of the
apparatus for CT image denoising based on deep learning (the
present apparatus 10) may be identically applied to the method for
CT image denoising based on deep learning, even though they are
omitted hereinafter.
[0112] Referring to FIG. 2, in a process S11 of the method for CT
image denoising based on deep learning according to an embodiment
of the present disclosure, the extraction unit 11 may extract
examination information from an input CT image.
[0113] Then, in a process S12, the selection unit 12 may select at
least one deep learning model corresponding to the examination
information extracted in the process S11 from multiple previously
trained deep learning models.
[0114] Then, in a process S13, the output unit 13 may output a CT
image denoised from the input CT image by feeding the input CT
image into at least one deep learning model selected in the process
S12.
[0115] Herein, in the process S13, the output unit 13 may feed the
input CT image into the at least one deep learning model selected
in the process S12 and thus force at least one deep learning model
selected in the process S12 to extract a noise-component CT image
from the input CT image, and output a CT image denoised from the
input CT image by multiplying the extracted noise-component CT
image with a predetermined value (e.g., a predetermined ratio) and
subtracting the result from the input CT image.
[0116] Meanwhile, the method for CT image denoising based on deep
learning according to an embodiment of the present disclosure may
further include generating and training multiple training-target
deep learning models before the process S11. Herein, the multiple
training-target deep learning models trained through the training
process may refer to the multiple previously trained deep learning
models considered in the process S12.
[0117] Herein, the method for training the multiple training-target
deep learning models, i.e., the method for training a deep learning
model, will be described with reference to FIG. 3 for easier
understanding.
[0118] In the descriptions above, the processes S11 to S13 may be
divided into additional processes or combined into fewer processes
depending on an embodiment. In addition, some of the processes may
be omitted and the sequence of the processes may be changed if
necessary.
[0119] FIG. 3 is a diagram showing a schematic operation flow
example of the method for training a deep learning model for CT
image denoising based on deep learning through an apparatus for CT
image denoising based on deep learning according to an embodiment
of the present disclosure. That is, FIG. 3 is a diagram showing a
schematic operation flow example of the method for training a
training-target deep learning model considered in the present
disclosure.
[0120] The method for training a deep learning model for CT image
denoising based on deep learning illustrated in FIG. 3 can be
performed by the above-described apparatus for CT image denoising
based on deep learning (the present apparatus 10). Accordingly, the
descriptions of the apparatus for CT image denoising based on deep
learning (the present apparatus 10) may be identically applied to
the method for training a deep learning model for CT image
denoising based on deep learning, even though they are omitted
hereinafter.
[0121] Referring to FIG. 3, in a process S21, the training unit 14
may input the CT image into a CT image noise simulator to train
deep learning models included in the present apparatus 10 before
the extraction unit 11 extracts the examination information from
the input CT image in the above-described process S11 of FIG.
2.
[0122] That is, in the process S21, the training unit 14 may
provide the CT image noise simulator with the input CT image as
input.
[0123] To be specific, in the process S21, the CT image noise
simulator may generate a synthetic sinogram by inputting an image
(original image) of a first training CT data set. Further, the CT
image noise simulator may generate a noise-component synthetic
sinogram by applying a noise model having a predetermined level
(desired level) of noise to the synthetic sinogram. Furthermore,
the CT image noise simulator may generate a noise-component CT
image based on the generated noise-component synthetic sinogram.
Moreover, the CT image noise simulator may generate a synthetic
high noise CT image by adding the generated noise-component CT
image to the original CT image.
[0124] Accordingly, in the process S21, the training unit 14 may
generate a second training CT data set to which noises of multiple
predetermined levels are added by applying the CT image noise
simulator to the first training CT data set before the extraction
unit 11 extracts the examination information from the input CT
image.
[0125] That is, in the process S21, the training unit 14 may
receive the first CT data set as input and generate a second
training CT data set with noises of multiple predetermined levels
by using the CT image noise simulator.
[0126] Herein, the second training CT data set may be composed of a
pair of synthetic high noise CT image set for respective noise
levels and synthetic noise-component CT image set for respective
noise levels obtained on the basis of the first training CT data
set.
[0127] Accordingly, the training unit 14 may train the
training-target deep learning models using a pair of the synthetic
high noise CT image set for respective noise levels and the
synthetic noise-component CT image set for respective noise levels
obtained by applying the CT image noise simulator to the first
training CT data set.
[0128] Then, in a process S22, the training unit 14 may extract
examination information from the second training CT data set and
group the second training CT data set into multiple groups
according to a predetermined rule.
[0129] Herein, in the process S22, the training unit 14 can
generate multiple groups of second training CT data set for
respective examination areas and/or respective scan properties by
grouping the second training CT data set into multiple groups.
[0130] Then, in a process S23, the training unit 14 may generate
and train multiple training-target deep learning models
corresponding to the multiple groups of second training CT data set
which have been generated in the process S22. That is, in the
process S23, the training unit 14 may generate and train multiple
training-target deep learning models corresponding to the multiple
groups of second training CT data set grouped in the process
S22.
[0131] To be specific, in the process S23, the training unit 14 may
generate and train multiple training-target deep learning models
corresponding to the respective groups of second training CT data
set generated for the respective examination areas. Further, the
training unit 14 may generate and train training-target multiple
deep learning models corresponding to the respective groups of
second training CT data set generated for the respective scan
properties.
[0132] In the process 23, the training unit 14 may train each of
the multiple training-target deep learning models generated in the
process S23 using the corresponding group of the second training CT
data set generated in the process S22.
[0133] As a specific example, in the process S23, the training unit
14 may use each group of second training CT data set generated for
respective examination areas and/or respective scan properties in
the process S22 to train each of the training-target deep learning
models which have been generated for the respective examination
areas and/or respective scan properties in the process S23.
[0134] Further, in the process S23, the training unit 14 may feed
the synthetic high noise CT image set for the respective groups of
the second training CT data set in the S22 as input into the
corresponding training-target deep learning model and train the
training-target deep learning model repeatedly by groups to
minimize the difference between the synthetic noise-component CT
image and output from the training-target deep learning model in
order for the training-target deep learning model to have a
function of extracting a noise-component CT image from an input CT
image.
[0135] In other words, the training unit 14 may feed each synthetic
high noise CT image for the respective groups of the second
training data set as an input into the corresponding
training-target deep learning model sequentially and train the
training-target deep learning model repeatedly by groups to
minimize the difference between the synthetic noise-component CT
image and output from the training-target deep learning model.
[0136] In the process S23, the training unit 14 may train each of
the multiple training-target deep learning generated for the
respective the examination areas using the corresponding group of
the second training CT data set with the same examination area
repeatedly. Further, the training unit 14 may train each of the
training-target deep learning models generated for the respective
scan properties using the corresponding group of the second
training CT data set with the same scan property repeatedly.
[0137] Accordingly, the multiple training-target deep learning
models considered by the present apparatus 10 may include the
training-target deep learning models generated for the respective
examination areas and the training-target deep learning models
generated for the respective scan properties.
[0138] In the descriptions above, the processes S21 to S23 may be
divided into additional processes or combined into fewer processes
depending on an embodiment. In addition, some of the processes may
be omitted and the sequence of the processes may be changed if
necessary.
[0139] The method for CT image denoising based on deep learning and
the method for training a deep learning model for CT image
denoising based on deep learning according to an embodiment of the
present disclosure may be implemented in an executable program
command form by various computer means and be recorded in a
computer-readable storage medium. The computer-readable storage
medium may include a program command, a data file, and a data
structure individually or a combination thereof. The program
command recorded in the computer-readable storage medium may be
specially designed or configured for the present disclosure or may
be known to those skilled in a computer software field to be used.
Examples of the computer-readable storage medium include magnetic
media such as hard disk, floppy disk, or magnetic tape, optical
media such as CD-ROM or DVD, magneto-optical media such as
floptical disk, and a hardware device such as ROM, RAM, flash
memory specially configured to store and execute program commands.
Examples of the program command include a machine language code
created by a complier and a high-level language code executable by
a computer using an interpreter. The hardware device may be
configured to be operated as at least one software module to
perform an operation of the present disclosure, and vice versa.
[0140] Further, the above-described method for CT image denoising
based on deep learning and method for training a deep learning
model for CT image denoising based on deep learning may be
implemented as a computer program or application stored in a
storage medium and executed by a computer.
[0141] The above description of the present disclosure is provided
for the purpose of illustration, and it would be understood by
those skilled in the art that various changes and modifications may
be made without changing technical conception and essential
features of the present disclosure. Thus, it is clear that the
above-described embodiments are illustrative in all aspects and do
not limit the present disclosure. For example, each component
described to be of a single type can be implemented in a
distributed manner. Likewise, components described to be
distributed can be implemented in a combined manner.
[0142] The scope of the present disclosure is defined by the
following claims rather than by the detailed description of the
embodiment. It shall be understood that all modifications and
embodiments conceived from the meaning and scope of the claims and
their equivalents are included in the scope of the present
disclosure.
* * * * *